The ASTRO509-3 dark energy puzzle Probability Distributions I have had my results for a long time: but I do not yet know how I am to arrive at them. Johann Carl Friedrich Gauss 1777-1855 ASTR509 Jasper Wall Fall term 2013 1
The In the dark previous energy episode... puzzle We considered the (shaky) basis for defining probabilities (= `belief ); at the rules which define probability theory, and at conditional probability. This led us to Bayes theorem: We considered how the terms broke down into posterior distribution (what we want to know), the likelihood (relative probability of the data given the model), the prior distribution, and the normalization term. We demonstrated Bayes in action with simple examples, urns and supernovae. We discussed how it is that Bayes theorem allows us to make inferences from the data, rather than compute the data we would get if we happened to know all relevant information. i.e. we showed that inverse problems are the real problems. We started to work on priors, used them in simple examples, noting that the prior really means what we know apart from the data. We discussed characterization of the posterior distribution. 2
The Probability dark energy distributions puzzle Example: toss four `fair' coins. The probabilities: - no heads = (1/2) 4 ; - one head = 4 x (1/2) 4 ; - two heads = 6 x (1/2) 4, etc. - sum of the possibilities for getting 0 heads to 4 heads = 1.0. If x is the number of heads (0,1,2,3,4), we have a set of probabilities prob(x) = (1/16,1/4,3/8,1/4,1/16); we have a probability distribution describing expectation of occurrence of event x. This probability distribution is discrete; there is a discrete set of outcomes and so a discrete set of probabilities for those outcomes. 3
The Probability dark energy distributions puzzle 2 In other words we have a mapping between the outcomes of the experiment and a set of integers. - sometimes the set of outcomes maps onto real numbers instead; here we discretize the range of real numbers into little ranges within which we assume the probability does not change. - If x is the real number that indexes outcomes, we associate with it a probability density f(x); the probability that we will get a number `near x, say within a tiny range δx, is prob(x) δx. - we loosely refer to probability distributions' with discrete outcomes or not. 4
The Probability dark energy distributions puzzle 3 Formally: if x is a continuous random variable, then f(x) is its probability density function, commonly termed probability distribution, when 1. Probability 2., and 3. f(x) is a single-valued non-negative number for all real x. The corresponding cumulative distribution function is Probability distributions and distribution functions may be similarly defined for sets of discrete values of x. Distributions may be multivariate, functions of more than one variable. 5
The Probability dark energy distributions puzzle 4 Quantifiers - location (where is the `centre'?) - dispersion (what is the `spread?) These quantifiers can be given by the first two moments of the distributions: Other moments, particularly the third moment ( skewness ) can play a prominent role; but these two are far the most important. 6
The Probability dark energy distributions puzzle 5 There are probability distributions we can calculate resulting from ideal experiments, outcomes or combinations of these. The best-known are the UNIFORM, BINOMIAL, POISSON and GAUSSIAN (or NORMAL) distributions, and these have a bunch of hangers-on 7
The dark common energy probability puzzle density functions 8
The dark Binomial energy Distribution puzzle There are two outcomes - `success' or `failure'. This common distribution gives the chance of n successes in N trials, with the probability of a success at each trial ρ, and successive trials are independent. This probability is Bernoulli, Johann, 1667-1748 The leading term, the combinatorial coefficient, gives the number of distinct ways of choosing n items out of N: 9
This coefficient can be derived as follows. There are N equivalent ways of arranging the N trials. However there are n permutations of the successes, and (N-n) permutations of the failures, which correspond to the same result namely, exactly n successes, arrangement unspecified. Since we require not just n successes (probability ρ n ) but exactly n successes, we need exactly N-n failures, probability (1- ρ) (N-n) as well. The binomial distribution follows from this argument. The binomial distribution has a mean value given by and a variance or mean square value of The dark Binomial energy Distribution puzzle 2 10
The dark Binomial energy Distribution puzzle - Example In a sample of 100 galaxy clusters selected by automatic techniques,10 contain a dominant central galaxy. We plan to check a different sample of 30 clusters, now selected by X-ray emission. How many of these clusters do we expect to have a dominant central galaxy? If we assume that the 10 per cent probability holds for the X-ray sample, then the chance of getting n dominant central galaxies is E.g.the chance of getting 10 is about 1%; if we found this many we would be suspicious that the X-ray cluster population differed from the general population. Suppose we made these observations and did find 10 centrally-dominated clusters. What can we do with this information? A Bayesian calculation that parallels the supernova example Assuming the X-ray galaxies are a homogeneous set, we can deduce the probability distribution for the fraction of these galaxies that have a dominant central galaxy. A relevant prior would be the results for the original larger survey. 11
The Binomial dark example, energy puzzle continued... The posterior probability distribution for the observation that 10/30 X-ray- selected clusters are centrally-dominated. The dark black line uses a uniform prior distribution for this fraction; the dashed line uses the prior derived from an assumed previous sample in which 10 out of 100 clusters had dominant central members. The light curve shows the distribution for this earlier sample. The figure makes clear that the data are not really sufficient to alter our prior very much. For example, there is only a 10 per cent chance that the centrally-dominant fraction exceeds even 0.2; the possibility of it being as high as 33% is completely negligible. Our X-ray clusters differ markedly from the general population. 12
The Poisson dark energy Distribution puzzle The Poisson distribution derives from the binomial in the limiting case of very rare events and a large number of trials, so that although ρ 0, N ρ (a finite value). Calling this finite mean value µ, the Poisson distribution is The variance of the Poisson distribution is also µ. Example : Village blacksmiths are/were occasionally kicked by the horse they were shoeing, say on average, 3 times per year. How often would they have good years with no kicks? How often would they have bad years, say 10 kicks? Poisson, Siméon-Denis, 1781-1840 13
The dark Poisson energy Distribution puzzle - Example 2 A familiar example of a process obeying Poisson statistics is the number of photons arriving during an integration. The probability of a photon arriving in a fixed interval of time is (often) small. The arrivals of successive photons are independent. Thus the conditions necessary for the Poisson distribution are met. Hence, if the integration over time t of photons arriving at a rate λ has a mean of µ = λt photons, then the fluctuation on this number will be σ = µ, because we know that the variance is µ. (In practice we usually only know the number of photons in a single exposure, rather than the mean number; obviously we can then only estimate the µ.) For photon-limited observations, such as CCD images or spectra, If we ``integrate" more, Thus Signal/Noise t, the sky-limited case. 14
The Poisson dark example energy puzzle 2, continued 15
The dark Gaussian energy (Normal) puzzle Distribution Both the Binomial and the Poisson distributions tend to the Gaussian distribution, large N in the case of the Binomial, large µ in the case of the Poisson. The (univariate) Gaussian (Normal) distribution is from which it is easy to show that the mean is µ and the variance is σ 2. 16
The dark Gaussian energy (Normal) puzzle Distribution 2 For the binomial when the sample size is very large, the discrete distribution tends to a continuous probability density in which the mean µ = N p and variance σ 2 = N p (1-p) are still given by the parent formulae for the binomial distribution. Here is an instance of the discrete changing to the continuous distribution: in this approximation we can treat n as a continuous variable (because n changes by one unit at a time, being an integer => the fractional change 1/n is small). 17
The dark Gaussian energy Distribution puzzle 3 18
The dark Central energy Limit puzzle Theorem The true importance of the Gaussian distribution and its dominant position in experimental science, stems from the Central Limit Theorem. A non-rigorous statement of this is as follows. THIS MAY BE THE MOST REMARKABLE THEOREM EVER It says that averaging will produce a Gaussian distribution of results - no matter the shape of distribution from which the sample is drawn. Eyeball integration counts Errors on averaged samples will always look `Gaussian'. The Central Limit Theorem shapes our entire view of experimentation. => error language of sigmas, describing tails of Gaussian distributions. 19
The Central dark energy Limit puzzle Theorem - Example 1 200 values drawn from exponential distribution with cutoff; averages of 1, 2, 4, 16 20
The dark Central energy Limit puzzle Theorem - Example 2 21
The Gaussian dark tails energy puzzle 22
The dark Power energy Law Distribution puzzle N(>L) = K L γ+1, integral form, N > L, or dn = (γ+1) K L γ dl, differential form, dn in dl not formally a probability distribution because = ; normally there are physical bounds so that it works scale-free: f(cl) = (cl) γ = Const x (L) γ = Const x f(l) Steep power laws -4 < γ < 0 pop up in astronomy frequently Criticality: earthquakes, stock-market fluctuations, forest fires, sub-networks on the internet, sand-piles..salpeter Mass Function, source counts (number-magnitude counts), primordial fluctuation spectrum. 23
The dark Power energy Law Distribution puzzle 2 Why do disasters occur? There are many pitfalls and combinations of pitfalls 1. It is totally different in character from binomial - Poisson - Gaussian Characterizing by means or variances completely misleading. The Central Limit Theorem fails us badly. 2. The index - differential or integral? - binning: uniform or on a ΔlogL scale? n.b. d(logl) = Const x L -1 dl 24
The dark Power energy Law Distribution puzzle - Example 25